A blueprint for building an agentic harness for your data science team, a structured context layer that turns any AI coding assistant (Cursor, Claude Code, Copilot, etc.) into a competent data science teammate.
This repository also configures its own agents. See what data-science-agent-toolkit tells them →
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Skill "<kebab-case-name>" from nandita7krishnan/data-science-agent-toolkit, covering inputs, outputs, procedure, big-data / safety guards and worked example.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Turn a fuzzy stakeholder question into a structured plan — clarifying questions, hypotheses, metrics, data sources, method, and deliverable shape — before any query runs.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Turn notebook output into a structured summary tailored to the audience — exec, Slack, email, or slide bullets — with the headline first.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Plan a safe backfill — scope, batching, idempotency precheck, cost estimate, monitoring, rollback path — output a runbook before any execution.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Build a cohort retention or conversion analysis — define the cohort, the time windows, and produce the retention triangle plus chart-ready output.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Consumer-side ad-hoc data audit — duplicates, nulls, schema drift, freshness, row-count sanity — before trusting a table for an analysis.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Profile any table — row counts, nulls, distributions, ranges, anomalies — sample-aware for big data, with output as a structured markdown summary.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Build a feature pipeline for tabular ML — encode, scale, impute, date-feature, assemble with ColumnTransformer, and run a leakage check before fitting anything.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Build a step-by-step conversion funnel — drop-off attribution, time-to-convert, segmented breakouts — from event data.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Write or modify an incremental SQL/dbt model — pick the right strategy by grain/scale/platform, configure uniquekey/cluster/eventtime, handle late-arriving data and idempotency.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Calculate a metric using the team's exact canonical definition from docs/metric-definitions.md — never improvise.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Score a trained model honestly — pick the right metric, compare to a naive baseline, report bootstrap CIs, plot residuals/calibration, and break out by slices.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Fresh-eyes review of an analysis notebook — hidden assumptions, missing CIs, sample-size issues, p-hacking risk, missing seed, cell ordering, missing data lineage — outputs ready-to-paste review comments.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Detect and mask PII in query output, logs, or notebook cells before sharing — applies the right transformation per data class from docs/data-classification.md.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Investigate a failed scheduled job — pull run status + error trace, classify transient vs code bug, check recent commits, output remediation plan.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Produce publication-quality charts that respect the data and the reader — chart-type selection, visual hierarchy, color, annotation, and the team's palette.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Pre-execution SQL review — partition filters, SELECT , cross joins, scan estimates, deterministic ordering, type coercion — surface issues before the query runs.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
When a metric moves unexpectedly, decompose it — by dimension, cohort, change-point — to localize the cause before proposing a hypothesis.
nandita7krishnan/data-science-agent-toolkit
Skill Claude CodeCodex
Meta-skill — scan code + DESCRIBE TABLE → draft a docs/schema/ .md from the template, flagging what only a human can know.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: